Repository Public
1{2 "id": "htr-tiny-0.3.0",3 "name": "HTR-Tiny 0.3.0",4 "family": "htr",5 "arch": "m1",6 "version": "0.3.0",7 "n_params": 33538945,8 "base_model": null,9 "licence": "Apache-2.0 (code and weights); non-commercial research use where CC BY-NC sources were included — see LICENSE-MODEL.txt",10 "default": false,11 "speed_hint": "≈0.3 s per paragraph (greedy or beam 3, Apple GPU)",12 "max_len": 640,13 "text_meta_prefix": false,14 "metrics": {15 "reference": "Differences in chrF++ versus the identity baseline (B0: return the input unchanged) on the Medium Haiku rewrite→original pairs; test = document-safe test split, held-out domain/period = generalisation splits of the paper. Exploratory unless marked confirmatory. Number recall and placeholder exactness are preservation rates on the test split; Pangram = share of reconstructions the Pangram classifier labels Human (54 passages).",16 "val_chrf_delta": null,17 "test_chrf_delta": 0.44,18 "test_chrf_ci": [19 0.3,20 0.5821 ],22 "test_chrf_pm": null,23 "test_confirmatory": true,24 "heldout_domain_chrf_delta": -0.1,25 "heldout_period_chrf_delta": -0.53,26 "number_recall": 0.991,27 "placeholders_exact": 0.992,28 "pangram_human": 54,29 "pangram_n": 54,30 "pangram_mixed": 0,31 "pangram": {32 "detector": "Pangram 4.1",33 "date": "2026-11-10",34 "n": 54,35 "human_share": 1.0,36 "flagged_share": 0.0,37 "mean_ai_likelihood": 0.0,38 "mean_ai_likelihood_ci": [39 0.0,40 0.041 ],42 "labels": {43 "human": 54,44 "mixed": 0,45 "ai": 046 },47 "reference": {48 "originals": {49 "human_share": 1.0,50 "flagged_share": 0.0,51 "mean_ai_likelihood": 0.052 },53 "rewrites": {54 "human_share": 0.537,55 "flagged_share": 0.463,56 "mean_ai_likelihood": 0.187557 }58 },59 "source": "artifacts/metrics/pangram_test/pangram_summary.md (system htr-m1-tiny-s3; flagged = labelled Mixed or AI; mean = word-weighted ai_assistance_score, 95 % document-clustered bootstrap CI)",60 "note": "This measures one commercial detector's behaviour on a 54-passage sample. A low AI-likelihood is not evidence of human authorship, and HTR must not be used to evade AI-text detection or for integrity decisions."61 }62 },63 "notes": [64 "The paper's confirmatory model M* (selected by the frozen validation rule, seed 3).",65 "Architecture m1 does not condition on the context fields (discipline, era, section); they are recorded with the run only."66 ],67 "trained_on": "Medium stage of WP-AI-0002: denoising pretraining on 258 264 rights-verified pre-2018 passages (PMC Open Access, CC BY economics and finance journals, arXiv, public-domain books), then fine-tuning on 72 698 validated Haiku rewrite → original pairs of 14 000 passages (32k SentencePiece).",68 "exported_at": "2026-11-10T06:45:15Z",69 "hf_repo": "spboucher04/htr-tiny",70 "hf_url": "https://huggingface.co/spboucher04/htr-tiny"71}